An Effective Hybrid Approach Based on Machine Learning Techniques for Auto-Translation: Japanese to English

Abstract

In recent years machine learning techniques have been able to perform tasks previously thought impossible or impractical such as image classification and natural language translation, as such this allows for the automation of tasks previously thought only possible by humans. This research work aims to test a naïve post processing grammar correction method using a Long Short Term Memory neural network to rearrange translated sentences from Subject Object Verb to Subject Verb Object. Here machine learning based techniques are used to successfully translate works in an automated fashion rather than manually and post processing translations to increase sentiment and grammar accuracy. The implementation of the proposed methodology uses a bounding box object detection model, optical character recognition model and a natural language processing model to fully translate manga without human intervention. The grammar correction experimentation tries to fix a common problem when machines translate between two natural languages that use different ordering, in this case from Japanese Subject Object Verb to English Subject Verb Object. For this experimentation 2 sequence to sequence Long Short Term Memory neural networks were developed, a character level and a word level model using word embedding to reorder English sentences from Subject Object Verb to Subject Verb Object. The results showed that the methodology works in practice and can automate the translation process successfully.